Invariant Object Recognition

1. A computer-implemented method of object recognition, the method comprising: training simulated neural circuitry using sequences of images representing moving objects, such that the simulated neural circuitry recognizes objects by having the presence of lower level object features that occurred in temporal sequence in the images representing moving objects trigger the activation of higher level object representations; receiving an image of an object, the image including lower level object features; and recognizing the object using the trained simulated neural circuitry, wherein the trained simulated neural circuitry activates a higher level representation of the object in response to the lower level object features from the image, wherein the simulated neural circuitry includes simulated neuronal elements with local nonlinear integrative properties.

2. The method of claim 1 , wherein the simulated neural circuitry includes elements having multiple input branches and training the simulated neural circuitry comprises: initializing connections between elements of the neural circuitry; and exposing the simulated neural circuitry to sequences of images representing moving objects, wherein the simulated neural circuitry is configured such that, as images of the objects simulate movement of the objects through a visual field of the simulated neural circuitry, consecutively active connections at a common input branch will be strengthened.

3. The method of claim 2 , wherein the simulated neural circuitry is configured such that strengthening of connections jointly strengthens inputs which co-occur within a specified window in time.

4. The method of claim 3 , wherein activity at one connection to the input branch causes effects on the input branch that remain active for a period of time spanning the duration of more than a single image in the sequence of images.

5. The method of claim 1 , wherein the images represent movement of the objects between different starting and ending positions.

6. The method of claim 1 , wherein the images represent rotation of the objects about one or more axes.

7. The method of claim 1 , wherein the images represent dilation of the objects.

8. The method of claim 1 , wherein a simulated neuronal element includes multiple input branches, one or more of which have local nonlinear integrative properties.

9. The method of claim 1 , wherein the simulated neural circuitry is configured such that connections between different simulated neuronal elements have different temporal properties, with some weakening and some strengthening with repeated input of a given frequency.

10. The method of claim 8 , wherein the simulated neural circuitry includes simulated neuronal elements configured to homeostatically regulate activity in the simulated neural circuitry.

11. The method of claim 1 , wherein the simulated neural circuitry functions as a self-organizing content-addressable memory.

12. The method of claim 1 , wherein: the simulated neural circuitry includes a simulated neuronal element represented as including multiple branches, a particular branch is configured to activate when the particular branch receives input that exceeds a first threshold, and the simulated neuronal element is configured to activate when a number of the simulated neuronal element's branches that are activated exceeds a second threshold.

13. The method of claim 12 , wherein the simulated neural circuitry includes neuronal elements of different types.

14. The method of claim 12 , wherein the simulated neural circuitry includes multiple processing areas, each of which includes multiple neuronal elements.

15. The method of claim 14 , wherein the processing areas are arranged in a hierarchical series from a lowest processing area to a highest processing area and feed-forward is provided from outputs of the neuronal elements of one processing area to inputs of neuronal elements of a higher processing area.

16. The method of claim 15 , wherein each higher processing area issues feedback to a lower processing area or areas.

17. The method of claim 14 , wherein the neuronal elements include dendritic trees.

18. A computer program embodied on a non-transitory computer-readable medium, the computer program including instructions that, when executed, cause a computer to: train simulated neural circuitry using sequences of images representing moving objects, such that the simulated neural circuitry recognizes objects by having the presence of lower level object features that occurred in temporal sequence in the images representing moving objects trigger the activation of higher level object representations; receive an image of an object, the image including lower level object features; and recognize the object using the trained simulated neural circuitry, wherein the trained simulated neural circuitry activates a higher level representation of the object in response to the lower level object features from the image, wherein the simulated neural circuitry includes simulated neuronal elements with local nonlinear integrative properties.

19. The computer program of claim 18 wherein the simulated neural circuitry includes elements having multiple input branches, and the instructions that cause a computer to train the simulated neural circuitry comprise instructions that, when executed, cause a computer to: initialize connections to input branches of the elements of the neural circuitry; and expose the simulated neural circuitry to sequences of images representing moving objects, wherein the computer program further comprises instructions that, when executed, cause a computer to strengthen consecutively active connections at a common input branch as images of the objects simulate movement of the objects through a visual field of the simulated neural circuitry.

20. The computer program of claim of claim 19 , wherein the instructions that, when executed, cause a computer to strengthen consecutively active connections at a common input branch include instructions that, when executed, cause a computer to jointly strengthen inputs which co-occur within a specified window in time.

21. The computer program of claim 18 , wherein a simulated neuronal element includes multiple input branches, one or more of which have local nonlinear integrative properties.

22. The computer program of claim 18 , wherein: the simulated neural circuitry is configured such that connections between different simulated neuronal elements have different temporal properties, and the computer program includes instructions that, when executed, cause a computer to weaken some connections and strengthen some connections in response to repeated input of a given frequency.

23. The computer program of claim 18 , wherein: the simulated neural circuitry includes a simulated neuronal element represented as including multiple branches, and the computer program includes instructions that, when executed, cause a computer to: activate a particular branch when the particular branch receives input that exceeds a first threshold, and activate the simulated neuronal element when a number of the simulated neuronal element's branches that are activated exceeds a second threshold.

24. The computer program of claim 18 wherein: the simulated neural circuitry includes processing areas arranged in a hierarchical series from a lowest processing area to a highest processing area, and the computer program includes instructions that, when executed, cause a computer to provide feed-forward from outputs of the neuronal elements of one processing area to inputs of neuronal elements of a higher processing area.

25. The computer program of claim 24 , wherein the computer program includes instructions that, when executed, cause a computer to issue feedback from each higher processing area to a lower processing area or areas.

26. An object recognition system comprising: means for receiving images; and simulated neural circuitry implemented by a processor or electronic circuitry for recognizing objects in received images, wherein the simulated neural circuitry is configured to: during a training process in which sequences of images representing moving objects are exposed to the neural circuit, learn to recognize the objects by having the presence of lower level object features that occur in temporal sequence in the images representing the moving objects trigger the activation of higher level object representations; and after the training process, recognize an object in a received image that includes lower level object features based on activating a higher level representation of the object in response to the lower level object features from the image, wherein the simulated neural circuitry includes simulated neuronal elements with local nonlinear integrative properties.

27. The object recognition system of claim 26 , wherein: the simulated neural circuitry includes elements having multiple input branches; and during the training process, the simulated neural circuitry is further configured to: initialize connections to input branches of elements of the neural circuitry; and strengthen consecutively active connections at a common input branch.

28. The object recognition system of claim 27 , wherein the simulated neural circuitry is configured such that strengthening of connections jointly strengthens inputs which co-occur within a specified window in time.

29. The object recognition system of claim 26 , wherein a simulated neuronal element includes multiple input branches, one or more of which have local nonlinear integrative properties.

30. The object recognition system of claim 26 , wherein the simulated neural circuitry is configured such that connections between different simulated neuronal elements have different temporal properties, with some being configured to weaken and some being configured to strengthen with repeated input of a given frequency.

31. The object recognition system of claim 26 , wherein the simulated neural circuitry functions as a self-organizing content-addressable memory.

32. The object recognition system of claim 26 , wherein: the simulated neural circuitry includes a simulated neuronal element represented as including multiple branches, a particular branch is configured to activate when the particular branch receives input that exceeds a first threshold, and the simulated neuronal element is configured to activate when a number of the simulated neuronal element's branches that are activated exceeds a second threshold.

33. The object recognition system of claim 26 , wherein: the simulated neural circuit includes multiple processing areas having multiple neuronal elements, the processing areas being arranged in a hierarchical series from a lowest processing area to a highest processing area; and the simulated neural circuit is configured to provide feed-forward from outputs of the neuronal elements of one processing area to inputs of neuronal elements of a higher processing area.

34. The object recognition system of claim 33 , wherein each higher processing area is configured to issue feedback to a lower processing area or areas.

35. The object recognition system of claim 33 , wherein the neuronal elements include dendritic trees.

36. A computer-implemented method for training simulated neural circuitry having at least a first processing area and a second processing area to achieve invariant object recognition, wherein the first processing area and the second processing area include simulated neuronal elements, the method comprising: initializing connections between simulated neuronal elements of the first processing area and simulated neuronal elements of the second processing area; receiving, at the simulated neural circuitry, a series of different images of an object; in response to receiving the different images of the object, generating initial encoded representations of the object in the first processing area by activating one or more of the simulated neuronal elements of the first processing area for each received image of the object; transmitting the initial encoded representations of the object from the first processing area to the second processing area by transmitting signals from active simulated neuronal elements of the first processing area to the simulated neuronal elements of the second processing area to which the active simulated neuronal elements of the first area are connected; and strengthening connections to an individual simulated neuronal element of the second processing area when the connections to the individual simulated neuronal element of the second processing area are consecutively active.

37. The method of claim 36 , wherein strengthening connections to an individual simulated neuronal element of the second processing area when the connections to the individual simulated neuronal element of the second processing area are consecutively active includes strengthening connections to the individual simulated neuronal element of the second processing area when the connections to the individual simulated neuronal element of the second processing area are active within a specified window of time.

38. A computer-implemented method for recognizing an object in an image comprising: receiving an input image of an object at simulated neural circuitry that includes a hierarchy of processing areas and that has been trained to recognize one or more objects; in response to receiving the input image of the object, generating an initial encoded representation of the object; transforming the initial encoded representation of the object into an invariant output representation of the object by propagating the initial encoded representation of the object through the hierarchy of processing areas, wherein representations of the object generated in each succeeding processing area are increasingly less variant than the initial encoded representation; and recognizing the object based on the invariant output representation of the object, wherein the simulated neural circuitry includes simulated neuronal elements with local nonlinear integrative properties.

39. The method of claim 38 , wherein a simulated neuronal element includes multiple input branches, one or more of which have local nonlinear integrative properties.

40. The method of claim 38 , wherein the simulated neural circuitry functions as a self-organizing content-addressable memory.

41. The method of claim 38 , wherein: the simulated neural circuitry includes a simulated neuronal element represented as including multiple branches, a particular branch is configured to activate when the particular branch receives input that exceeds a first threshold, and the simulated neuronal element is configured to activate when a number of the simulated neuronal element's branches that are activated exceeds a second threshold.

42. A computer-implemented method for searching for images of an object, the method comprising: receiving a sequence of images of an object at simulated neural circuitry; training the simulated neural circuitry using the sequence of images to generate an encoded representation of the object, wherein the encoded representation of the object includes a higher level feature triggered by recognizing a temporal sequence of lower level features of the object during the training; searching a corpus of images by supplying each image of the corpus to the simulated neural circuitry; and designating images from the corpus of images as being images of the object when an output of the simulated neural circuitry produced as a result of an image being supplied to the simulated neural circuitry matches the encoded representation of the object to within a predetermined threshold, wherein the simulated neural circuitry includes simulated neuronal elements with local nonlinear integrative properties.

43. The method of claim 42 , wherein the sequence of images comprises a video of the object.

44. The method of claim 43 , wherein a simulated neuronal element includes multiple input branches, one or more of which have local nonlinear integrative properties.

45. The method of claim 43 , wherein the simulated neural circuitry includes simulated neuronal elements configured to regulate excitation levels and synaptic connection strengths.

46. The method of claim 43 , wherein the simulated neural circuitry functions as a self-organizing content-addressable memory.

47. The method of claim 43 , wherein: the simulated neural circuitry includes a simulated neuronal element represented as including multiple branches, a particular branch is configured to activate when the particular branch receives input that exceeds a first threshold, and the simulated neuronal element is configured to activate when a number of the simulated neuronal element's branches that are activated exceeds a second threshold.